Machine learning

• US 20050105712A1
• Filed: 09/10/2004
• Published: 05/19/2005
• Est. Priority Date: 02/11/2003
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Williams, David R., Hill, Jeffrey
• Granted Patent: US 7,606,714 B2
• Time in Patent Office: Days
• US Class Current: 379/265.20
• CPC Class Codes: G06F 16/3344 using natural language analysis
• CPC Class Codes: G10L 13/027 Concept to speech synthesisers; Generation of natural phrases from machine-based concepts generation of parameters for speech synthesis out of text G10L13/08
• CPC Class Codes: G10L 15/1822 Parsing for meaning understanding
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue
• CPC Class Codes: H04M 3/493 Interactive information services, e.g. directory enquiries ; Arrangements therefor, e.g. interactive voice response [IVR] systems or voice portals
• CPC Class Codes: H04M 3/5166 in combination with interactive voice response systems or voice portals, e.g. as front-ends

Abstract

An automated response system (e.g., an automated voice response system) may employ learning strategies to develop or improve automated response capabilities. Learning strategies may include using communications(e.g., utterances, text messages, etc.) of one party in a conversation (e.g., a customer service agent) to identify and categorize communications of another party in the conversation (e.g., a caller). Classifiers can be build from the categorized communications. Classifiers can be used to identify common communications patterns of a party in a conversation (e.g., an agent). Learning strategies may also include selecting communications as learning opportunities to improve automated response capabilities based on selection criteria (e.g., selection criteria chosen to ensure that the system does not learn from unreliable or insignificant examples).

Claims

1. A method comprising:

receiving a set of conversations between a members of a first party type and a members of a second party type, wherein each of the conversations includes a communication of a member of the first party type and a communication of a member of the second party type that is responsive to the communication of the member of the first party type;

grouping the communications of members of the first party type into a first set of clusters;

grouping the responsive communications of members of the second party type into a second set of clusters based upon the grouping of the communications of members of the first party type; and

by machine, generating a set of second party type classifiers for one or more clusters in the second set of clusters.

2. The method of claim 1 wherein the communications comprises utterances.

3. The method of claim 1 wherein the communications comprises text messages.

4. The method of claim 1 wherein the communications of members of the first party type comprise communications of human customer service agents at a call center.

5. The method of claim 1 wherein the communications of members of the first party type comprise communications of software agents configured to communicate with humans who contact a call center.

6. The method of claim 1 wherein the communications of members of the second party comprise communications of humans who have contacted a call center.

7. The method of claim 1 wherein the classifiers comprise support vector machines.

8. The method of claim 1 wherein the classifiers comprise decision trees.

9. The method of claim 1 wherein communications of members of a first party type are grouped into a first set of clusters using a computer.

10. The method of claim 9 wherein grouping communications of members of a first party type into a first set of clusters comprises determining semantic features of the communications.

11. The method of claim 1 wherein grouping communications of members of the first party type into a first set of clusters is based on a meaning of the communications of members of the first party type.

12. The method of claim 1 further comprising:

by machine, generating a set of first party type classifiers for one or more clusters in the first set of clusters.

13. The method of claim 1 wherein grouping communications of members of the first party type into a first set of clusters comprises:

grouping communications corresponding to requests for information from members of the first party type into a first set of clusters.

14. The method of claim 13 wherein grouping responsive communications of members of the second party type into a second set of clusters based upon the grouping of the communications of members of the first party type comprises:

grouping communications of members of the second party type into groups corresponding to responses to the requests for information from members of the first party type.

15. The method of claim 13 wherein grouping responsive communications of members of the second party type into a second set of clusters based upon the grouping of the communications of members of the first party type comprises:

using the first party type classifiers to classify a communication of a member of the first party type into a cluster of the first party type;

grouping a communication of a member of the second party type that is subsequent to the classified communication of the member of the first party type into a cluster of the second party type that relates to the cluster of the first party type.

16. The method of claim 15 wherein the cluster of the first party type relates to a request for information made by a member of the first party type and the cluster of the second party type relates to a response to the request for information given by a member of the second party type.

17. The method of claim 1 further comprising:

receiving a second set of conversations between members of the first party type and members of the second party type, wherein each of the conversations includes a communication of a member of the first party type and a communication of a member of the second party type that is responsive to the communication of the member of the first party type;

applying classifiers to group the communications of members of the second party type;

by machine, regenerating a second party type classifiers for a cluster in the second set of clusters using data relating to the communications grouped in the cluster.

18. A method comprising:

by machine, applying a set of classifiers to categorize initiating communications that are part of conversations that also include responsive communications; and

by machine, using the categorized initiating communication to identify common communication patterns.

19. The method of claim 18 further comprising:

grouping conversations in a set of conversations.

20. The method of claim 19 further comprising:

associating an identified common communication pattern with a group of conversations.

21. The method of claim 20 wherein the conversations are grouped by subject matter of the conversation.

22. The method of claim 18 wherein the communications comprise utterances.

23. The method of claim 22 wherein the conversations include communications from agents associated with a customer service call center and communications from customers who contacted the call center.

24. The method of claim 23 wherein the agents are human agents.

25. The method of claim 23 further comprising:

grouping conversation in the set of conversations according to a reason for the customers'"'"' contact with the call center.

26. A method comprising:

by machine, applying classifiers to identify a set of classified communications made by a member of a first party type in a conversation that also includes responsive communications made by a member of a second party type; and

by machine, determining a subject matter of each of the conversations based on the set classified communications of the member of the first party type in a conversation.

27. The method of claim 26 wherein the classifiers classify the communications according to a representation of the concepts embodied in the communication.

28. The method of claim 27 wherein determining a subject matter of the conversation based on the combination of categorized communications of the member of the first party type comprises:

matching the set of categorized communications with a set of categorized communications associated with a conversation having a known subject matter.

29. The method of claim 26 wherein determining the subject matter of the conversation based on the set of classified communications is made independent of any particular ordering of the classified communications in the set.

30. The method of claim 26 wherein the communications comprise utterances.

31. The method of claim 26 wherein the communications of members of the first party type comprise communications of customer service agents at a call center.

32. The method of claim 26 wherein the sequence of categorized communications comprises a sequence of requests made by the customer service agent.

33. The method of claim 31 wherein the customer service agents comprise humans.

34. The method of claim 31 wherein the customer service agents comprise software configured to communicate with human callers.

35. The method of claim 26 wherein the set of classifiers comprise a support vector machine.

36. A computer-implemented method comprising:

receiving digital representations of conversations at least some of which comprise a series of communications between a person and an agent associated with a contact center; and

selecting a communication as a learning opportunity if one or more selection criteria are satisfied.

37. The method of claim 36 wherein the communications comprise utterances.

38. The method of claim 36 wherein the communications comprise text messages.

39. The method of claim 36 wherein the agent associated with the contact center comprises a software agent configured to communicate with the persons.

40. The method of claim 36 wherein the agents associated with the contact center comprises a human agent who communicates with the persons.

41. The method of claim 36 wherein the selection criteria comprises:

a requirement that a communication be followed by communication exchanges between the person and one or more agents.

42. The method of claim 36 wherein the selection criteria comprise:

a requirement that a communication be followed by a predefined number of successful subsequent communication exchanges between the person and one or more agents.

43. The method of claim 36 wherein the selection criteria comprises:

a requirement that a communication be included within a conversation in which the person responded positively to a satisfaction question posed by the system.

44. The method of claim 36 wherein the selection criteria comprise:

a requirement that a communication in a first conversation be confirmed by similar communications occurring in a number of other conversations.

45. The method of claim 44 wherein the selection criteria comprise:

a requirement that at least one of the conversations in which the similar communications took place includes an indication that a person responded positively to a satisfaction question posed by the system.

46. The method of claim 36 wherein at least some of the communications between the persons and agents include assist interactions in which a human agent selected a response to a person'"'"'s communication from a ranked list of proposed responses generated by the automated response system.

47. The method of claim 46 wherein the selection criteria comprise:

a requirement that a selected response in an assist interaction be ranked above a threshold.

48. The method of claim 46 wherein the selection criteria comprise:

a requirement that a selected response in an assist interaction be selected from a trusted human agent.

49. The method of claim 48 wherein a trusted human agent is an agent who has been employed at that contact center for more than a predetermined length of time.

50. The method of claim 36 wherein the selection criteria comprise:

a requirement that a communication not cause a set of classifiers built using the communication to misclassify communications that a previous set of classifiers had classified correctly.

51. The method of claim 36 further comprising:

receiving one or more of the selection criteria from a user.

52. The method of claim 51 further comprising:

providing a graphical user interface through which the selection criteria may be entered.

53. The method of claim 36 further comprising:

using a selected communication to generate a classifier.

54. The method of claim 53 wherein the classifier comprises a support vector machine.

55. The method of claim 36 further comprising:

using a selected communication to generate a language model for an automatic speech recognition engine.

56. The method of claim 36 further comprising:

using a selected communication to produce a finite state network.

57. The method of claim 37 further comprising:

performing speech recognition on an utterance selected as a learning opportunity.

58. The method of claim 57 further comprising:

prior to performing the speech recognition, determining whether to perform speech recognition on the selected utterance based on a confidence level of the meaning of the utterance associated with the digital representation of the communication.

59. A computer-implemented method comprising:

receiving a digital representation of a conversation that includes a series of utterances between a caller and an agent associated with a contact center; and

after receiving digital representation, selecting the utterance for transcription based on one or more selection criteria.

60. The method of claim 59 wherein the selection criteria comprises:

a requirement that a confidence level of a response by the automated voice response system be within a range of values.

61. The method of claim 60 further comprising:

receiving the range of values from a user through a graphical user interface.

62. The method of claim 59 wherein the selection criteria comprises:

a requirement that a confidence level of speech recognition process performed on the utterance during the conversation is within a range of values.

63. The method of claim 62 further comprising:

receiving the range of values from a user through a graphical user interface.

64. The method of claim 59 further comprising:

performing a computer-implemented speech recognition process on the selected utterance.

65. The method of claim 59 further comprising:

adding recognized words in the selected utterance to a vocabulary of words used by a speech recognition process used by the system to recognize utterances during conversations.

66. A method comprising:

based on an interaction between a person and a human agent associated with an automated response system in which the agent selected a response to a communication of a person from among responses proposed by the automated response system; and

selecting the communication as an example to train the automated response system.

67. The method of claim 66 wherein selecting the communication comprises:

selecting the communication based on a confidence level of the selected response.

68. The method of claim 66 further comprising:

using the selected communication to train a classifier.

69. The method of claim 66 further comprising:

adding the selected communication to a language model of a statistical language model automatic speech recognition process.

70. The method of claim 66 wherein selecting the communication comprises:

selecting the communication based on a level of trust of a human agent who selected the response.

71. The method of claim 66 wherein the communication comprises an utterance.

72. A method comprising:

by machine, identifying a communication between a person contacting an response system and a human agent; and

modifying the automated response system to respond to similar future communications from persons contacting the system.

73. The system of claim 72 wherein modifying the automated response system to respond to similar future communications from persons contacting the system comprises:

modifying a finite state transition network associated with the system.

74. A computer-implemented method comprising:

adding a communication to a set of training examples for a classifier in a concept recognition engine;

generating a new classifier using the set of training examples that includes the added communication; and

disregarding the new classifier based on a performance requirement for a new classifier.

75. The method of claim 74 wherein the performance requirement comprises:

a requirement that a new classifier correctly classify at least a predetermined number of other examples.

76. The method of claim 74 wherein the performance requirement comprises:

a requirement that a new classifier have a new definitive set of examples that is different from the definitive set of examples of the previous classifier by a predetermined amount.

77. The method comprising:

generating a set of classifiers for at least one cluster of responsive communications, the cluster being based on one or more clusters of initiating communications with which the responsive communications are associated within conversations.

78. The method of claim 77 wherein the initiating conversations are from a member of a first party type.

79. The method of claim 77 wherein the responsive communications are from a member of a second party type.

80. The method of claim 78 wherein the first party type comprises agents associated with a customer contact center.

81. The method of claim 79 wherein the second party type comprises customers who contacted a customer contact center.

82. The method of claim 77 further comprising receiving a set of conversations at least some of which include an initiating communication and an associated responsive communications.

83. The method of claim 77 wherein the cluster of responsive communication comprises responsive communications associated with an initiating communication.